ELSEVIER Journalof NeuroscienceMethods 63 (1995) 43-54

Tetrodes markedly improve the reliability and yield of multiple

single-unit isolation from multi-unit recordings in cat striate cortex
Charles M. Gray ‘7* , Pedro E. Maldonado a, Mathew Wilson b*l, Bruce McNaughton b
’ Center for Neuroscience, University of California, 1544 Newton Court, Davis, CA 95616, L’SA
h Diuision of Neural Systems, Memory and Aging. lJniL;ersity of Arizona, Tucson, AZ 85712, USA

Received21 December1994;revised 2 June 1995;accepted7 June 1995

Abstract

The majority of techniquesfor separatingmultiple single-unitspiketrains from a multi-unit recordingrely on the assumption that
different cells exhibit action potentialshaving uniqueamplitudesand waveforms.When this assumptionfails. due to the similarity of
spikeshapeamongdifferent cellsor to the presence of complexspikeswith decliningintra-burstamplitude,thesemethodsleadto errors
in classification.In an effort to avoid theseerrors,the stereotrode(McNaughtonet al., 1983)andlater the tetrode(O’Keefe andReece,
1993;Wilson andMcNaughton,1993)recordingtechniquesweredeveloped.Becausethe latter techniquehasheenappliedprimarily to
the hippocampus, we soughtto evaluateits performancein the neocortex.Multi-unit recordings,usingsingletetrodes.weremadeat 28
sitesin area17of 3 anesthetized cats.Neuronswereactivatedwith moving barsandsquarewave gratings.Singleunits were separated by
identificationof clustersin 2-D projectionsof either peak-to-peakamplitude,spikewidth, spike area,or the 1st versus2nd principal
componentsof the waveformsrecordedon eachchannel.Usingtetrodes,we recordeda total of 154singlecells (mean= 5.4, max = 9).
By cross-checking the performanceof the tetrodewith the stereotrodeandelectrode,we foundthat the bestof the 6 possiblestereotrode
pairsand the bestof 4 possibleelectrodesfrom eachtetrodeyielded 102(mean= 3.6, max = 7) and 95 (mean= 3.4, max = 6) cells,
respectively.Moreover,we foundthat the numberof cells isolatedat eachsiteby the tetrodewasgreaterthanthe stereotrode or electrode
in 16/28 and 28/28 cases,respectively.Thus, both stereotrodes, and particularlyelectrodes,often lumped2 or morecells in a single
clusterthat could be easily separatedby the tetrode. We concludethat tetroderecordingcurrently provides the best and most reliable
methodfor the isolationof multiplesingleunits in the neocortexusinga singleprobe.

@words; Multi-unitrecording; Visualcortex,area17;Spikedetection;

Principalcomponents
analysis

1. Introduction minimal insight into the dynamics of interacting popula-

tions of cells. Direct knowledge of cortical network dy-
Physiological investigations of neocortical processing namicscan most effectively be obtained by recording from
typically seek to relate the discharge characteristics of a at least two and preferably many more neurons simultane-
single well-isolated neuron to the properties of a sensory ously. Unfortunately, multiple, single-neuron recording in
stimulus, a motor response, or a behavioral task. By mammalian cortex is fraught with a number of technical
repeating this process over many neurons recorded at problems. Chief among theseare the difficulties associated
different times, it is often possibleto infer some aspect of with recording stable single-unit activity from more than a
cortical network function from the cumulative behavior of few microelectrodes at the same time. Because of this and
the individual celis. This approach has met with great other limitations, many investigators have resorted to the
successover the past several decadesbut it has provided recording of multi-unit activity and the application of
waveform identification techniques to extract the spike
trains of several cells recorded by a single electrode. This
technique increasesthe number of cells that can be simul-
* Correspondingauthor. Tel.: (916) 754-5083; Internet: taneously recorded and reduces the requirement for main-
[email protected] or [email protected].
’ Present
address: Departmentof BrainandCognitiveScience,Mas- taining strict isolation of a single cell.
sachusetts
Instituteof Technology,
45 CarletonStreet,E25-236,Cam The principal methodsemployed to separatesingle units
bridge,MA 02139.USA.Internet:[email protected]. from a multi-unit recording rely on two assumptions:each

016%0270/95/$09.50
0 I995ElsevierScience
B.V. All rightsreserved
SSDI 0165.0270(95)00085-2
44 C.M. Gray et al. / Journal of Neuroscience Mefhods 63 (1995) 43-54

cell near the tip of a microelectrode exhibits action poten- identification (O’Keefe and Reece, 1993; Wilson and Mc-
tials whose waveforms remain constant in shape but differ Naughton, 1993). The performance of this method has
in amplitude and time course from each of the other been impressive. Yields of lo-15 well-isolated hippocam-
recorded cells. Satisfaction of these constraints has made pal cells on a single tetrode have not been uncommon
possible the use of such methods as amplitude discrimina- (Wilson and McNaughton, 1993). Because of these in-
tion, template matching and principal components analysis creased yields and the resistance of the method to classifi-
(Abeles and Goldstein, 1977; for review see Wheeler and cation errors, we became interested in testing the perfor-
Heetderks, 1982; WSrgiitter et al., 1986; Salganicoff et al., mance of tetrodes in the mammalian visual cortex. Here
1988; Kreiter et al., 1989; Jansen and Ter Maat, 1992). In we report the results of this analysis and compare the
a number of instances, the percentage of which is not performance of tetrodes to stereotrodes and electrodes.
accurately known, these assumptions are likely to be in-
valid, however. Many cells in the neocortex and other
structures, such as the hippocampus, fire action potentials 2. Surgery and anesthesia
in rapid bursts (Ranck, 1973; Connors et al., 1982; Mc-
Cormick et al., 1985; Agmon and Connors, 1989; Gray et Three adult male cats, weighing 4-5 kg, were used in
al., 1990). The shape and amplitude of the spikes within a the present study. On the day of the experiment the
burst change in characteristic ways. In cat visual cortex, animals were anesthetized with an intramuscular injection
for example, spike amplitude decreases and spike width of Ketamine (12 mg/kg) and Xylazine (1 mg/kg) and
increases during the course of a burst, resulting in action given Atropine (0.05 mg/kg s.c.> to reduce salivation. The
potentials of distinctly different waveform arising from the cephalic vein was cannulated and a continuous infusion of
same cell (Gray, 1992). These differences are often large Ringers containing 2.5% dextrose was given throughout
enough to lead to the identification of 2, or possibly even the experiment (4 ml/kg/h). Anesthesia was maintained
3, units from the activity of a single cell when single using halothane (0.6-l .5%) in a mixture of nitrous
electrode methods are employed. A second, and often oxide/oxygen (2: 1) while the animals were actively venti-
undetectable, error in classification can result when adja- lated using a Harvard respirator pump. The EKG, heart
cent neurons exhibit spikes that have very similar ampli- rate, rectal body temperature and expiratory CO, were
tude and time course. In these instances the activity from 2 continuously monitored, the latter three being maintained
or possibly more cells can be incorrectly lumped together. within the ranges of 140-180 bps, 37.5-39.0” and 3.5-
Such an error can only be detected when the calculation of 4.5%, respectively. The animal’s were mounted in a Kopf
the interspike interval histogram reveals the absence of a stereotaxic and a small craniotomy was made in one
well-defined refractory period. Such a control may not be hemisphere over the representation of the area centralis of
entirely accurate, however, if the cells fire at low rates or area 17. Following the surgery, the animals were paralyzed
at different times. Therefore, in those instances where the with pancuronium bromide (Pavulon) at a dose of 3 mg/kg
above assumptions are invalid, incorrect classification of (initial bolus i.v.> followed by a continuous infusion of 3
spike identity becomes very likely. mg/kg/h (i.v.1. The eyes were focused on the screen of a
In an effort to avoid these potentially serious classifica- computer monitor at a distance of 57 cm using the tapetal
tion errors, the stereotrode method was developed (Mc- reflection technique (Pettigrew et al., 1979) and an appro-
Naughton et al., 1983). This technique relies on the fact priate pair of gas permeable contact lenses. Following
that action potential amplitude is some declining function these procedures, a small opening was made in the dura
of distance between the electrode tip and the cell. There- matter, a single tetrode positioned just above the cortical
fore, if cellular activity is recorded from 2 closely spaced surface, and a 4% mixture of agar in Ringers solution
electrodes the relative amplitudes of the spikes on the 2 applied to the cortical surface to reduce pulsations. The
channels can be used as an additional criterion for spike assembly was then covered with molten bone wax.
separation. Moreover, if the spikes from a given cell
change shape and/or amplitude, as occurs during burst
firing, these changes are proportionally the same on the 2 3. Tetrode fabrication
channels, thus allowing accurate identification of these
cells. Not only does the method reduce errors in classifica- The tetrodes were fabricated according to previously
tion, but it also increases the number of cells that can be published methods (Wilson and McNaughton, 1993).
identified. In spite of these advantages, it has been recog- Briefly, 12 pm diameter insulated Nichrome wire (H.P.
nized that the stereotrode method may not be optimal Reid) was cut into lengths of approximately 25 cm and
(McNaughton et al., 1983). Two, and occasionally more, wound together (10 turns/inch) by grasping the wires with
cells can have action potentials of such similar properties a clip, suspending them with another clip attached to a
that they fail to be separated by the stereotrode. Thus an small metal rod, and rotating the suspended rod with a
extension of the technique termed the ‘tetrode’ has been magnetic stirring plate. The insulation of the wound wires
employed to increase the reliability and yield of single-unit was then melded together, while under static rotational
 C.M. Gray et al. / Journal of Neuroscience Merhods 63 11995143- 54 .a5

tension, by applying a small directed flame along roughly insulation and served not only to hold the wires closely
20 cm of the length of the suspended wires. The tempera- together but to also increase their rigidity. Tbe braided set
ture of the flame was insufficient to melt through the of wires were then cut in half and the unmelded ends

Fig. 1. The upper 6 graphs show scatter plots of the peak-to-peak amplitnde of all spikes recorded on 1 channel versus another for each of the 6 possible 2
channel comparisons obtained from a single tetrode. The channel combination is labeled in the upper right of each graph. The horizontal and vertical scale
bars indicate 13% and 17%. respectively, of the normalized peak amplitude across all channels. Note that each 2-D projection has a unique distribution of
clusters. In some propctions, such as l-3 and 2-4, 6 clusters can be discerned, while in other projections the cluster boundaries overlap to a greater degree
(l-2) or fewer clusters can be detected (3-4). The bottom 6 graphs of the figure show plots of 50 overlapping, unavemged waveforms f 1.1 ms in duration)
for each of the 6 identified single units (A-F) recorded on each channel (a-d) of the same tetrode. Note the clear differences in spike amplitude on
adjacent channels and how the ratios of these differences vary from cell to cell
46 C.M. Gray et al./ Journal of Neuroscience Methods 63 11995) 43-54

separated and the insulation removed by brief exposure to the Discovery software system (Datawave Technologies).
a small flame. During this procedure it was necessary to Spike waveforms of 1.l ms duration, centered around the
grasp the wires with a fine pair of forceps serving as a heat time of occurrence of a user-defined threshold crossing,
sink and to expose the wire to the flame for approximately were displayed continuously on-line and stored in a file
1 s. We then inserted the tetrode into a 30-ga stainless steel along with their associated time stamps, with a temporal
guide tube, allowed the tip to protrude approximately resolution of 0.1 ms. The threshold was defined separately
OS-l.0 cm, and fixed the wire to the tube with cyanoacry- for each channel, and spike waveforms were stored for all
late. The uninsulated ends of the 4 wires were inserted into 4 channels whenever any 1 of the 4 exceeded its threshold
a small Cpin connector, fixed with conductive epoxy, and value. During the course of data collection it was possible
the assembly sealed and fixed in place with dental acrylic, to display 2-dimensional (2-D) scatter plots of selected
In order to ease penetration into the cortex and to improve spike parameters on any 1 channel versus another. This
the recording characteristics, the tips of the tetrodes were allowed us to determine if a given multi-unit recording
beveled on 4 sides by briefly touching them to a fine was yielding clearly separable single units and, if the
grinding wheel at an angle of 30”. Each tip was then plated pattern of clusters was stable over time. If the cluster
with gold (gold cyanide SIFCO, code 3023) by passing DC pattern could be observed to drift over the course of
current until the impedance at 1 kHz was in the range of minutes we discontinued data collection until the pattern
0.5-1.0 MO. showed no further variation.

4. Data acquisition 5. Visual stimulation

The signals from each wire of the tetrode were ampli- Once stable unit recordings were obtained, we mapped
fied (lOk), bandpass filtered (0.6-6 kHz, 3 dB falloff), and the receptive field properties of the multi-unit activity for
digitized (30 kHz/chn, Data Translation DT2821) using both eyes using a mouse-controlled light bar presented on
an IBM compatible, 486-based, personal computer running a computer monitor at a distance of 57 cm from the eyes.

Fig. 2. Scatter plots of the dot product of each spike waveform with the 1st (abscissa)and 2nd (ordinate.)principal componentof an averagecortical
waveform. Each plot is derived from one of the four spike trains of the tetroderecordingshownin Fig. 1. The axes are both positive (up, right) and
negative (down, left) and the length of the scale bars represent 20% of the maximum across the plots. Data points lying close to the center of each display
represent waveforms that are close to the noise level. Clusters of points represent waveforms that exhibit similar principal components. This analysis is less
sensitive to variations in amplitude but very sensitive to differences in waveshape. Note that distinct clusters appear in each plot that vary in resolution on
different channels. For each channel, however, the number of clusters and the clarity of their separation is less than that observed with the 2-D projections
of peak-to-peak amplitude (compare Figs. 1 and 2).
 CM. Gray et al./Journal of Neuroscience Methods 63 (I9951 li-54 47

The mapping program allowed us to demarcate the bound- conventional spike extraction techniques using single elec-
aries of the receptive field and to define a region for trodes. To accomplish this, we subdivided the data ob-
subsequent visual stimulation. All visual stimuli were pre- tained at each recording site into 3 categories: 1 single
sented monocularly and consisted of either a drifting light tetrode, 6 stereotrode pairs, and 4 single electrodes. The
bar or square wave grating presented at the optimal orien- methods we employed for single-unit identification using
tation, direction, velocity and spatial frequency for the each category of data were similar to those published
multi-unit activity. We typically ran 20-40 trials for the previously (Abeles and Goldstein, 1977; McNaughton et
optimal stimulus and then advanced the tetrode to a new al., 1983; Wilson and McNaughton, 1993). For each wave-
location and repeated the process. Occasionally we also form in a data set, we computed the following parameters:
ran tuning curves for either direction of motion or spatial peak positive amplitude, peak negative amplitude, peak-
frequency. to-peak amplitude, spike width (defined as the time differ-
ence from peak negativity to peak positivity), the integral
of the spike waveform, and time of occurrence. In addi-
6. Spike extraction and single-unit identification tion, for the single electrode analysis, we employed the
method of principal components analysis to identify spike
The principal aim of our analysis was to compare the waveforms of similar shape (Abeles and Goldstein, 19771.
performance of tetrodes with that of stereotrodes and Briefly, we computed the 1st and 2nd principal compo-

1 2
TIME (set)
Fig. 3. Plots of the PSTHs for the 6 single units illustrated in Figs. 1 and 2. The cells were stimulated by monocular presentation of an optimally oriented
light bar passing over the receptive fields in the preferred direction of movement. The numbers to the right of each histogram indicate the number of spikes
per bin. The waveforms of 50 overlapping spikes recorded from each cell on each wire of the tebode are shown to the far right. These are the same
waveforms as those shown in Fig. 1. Note that cells 1-3 and 6 respond to the visual stimulus while cell 4 exhibits maintained firing and cell 5 only fires
occasionally.
48 C.M. Gray et al./Journal of Neuroscience Methods 63 (1995) 43-54

nents from a library of 80 spike waveforms recorded in products were plotted in a scatter diagram for each wave-
cortex. We then calculated the sum of the squares of the form recorded in a data set, where:
projections of the 1st and 2nd principal components for
each waveform according to the following relation: F, = E (e,i*wi) F2= E (ezi*wi) (2)
i= 1 i= 1

We further applied a threshold criteria to reject incoming

waveforms that were too close to the noise level. Thus,
any incoming waveform having a threshold value T not
C 4i C 4; exceeding 0.6 was plotted on the origin, where:
i= 1 i= 1

where w is the incoming waveform and e, and e, are the

1st and 2nd principal components, respectively. The value
cp(t) was computed over a range of 9 time lags, where the
time resolution was equivalent to the digitizing interval The principal result of this calculation is that similar spike
(i.e., l/30 kHz). At the time lag for which cp(t> reached a waveforms produce adjacent data points forming a cluster
maximum, the value of the 1st (F,) and 2nd (F,) dot (Abeles and Goldstein, 1977), while values not exceeding

11
22

1 2
TIME (see)
Fig. 4. PSTHs and associated waveforms for another set of 6 single units isolated from a different recording site. The cells were stimulated monocularly
with a high contrast, drifting, square-wave grating at the optimal orientation, spatial frequency and direction of motion. The calibrations of 4 and 2 refer to
cells l-3 and 4-6, respectively. Note that all 6 cells respond with strong modulation at the fundamental temporal frequency of the grating, indicative of
simple receptive field organization, but they segregate into 2 groups of 3 tiring in antiphase relation.
 CM Gray et al. /Journal of Neuroscience Methods 63 f 1995 ) 43- 54 J9

the threshold of 0.6 are excluded from the analysis. For biguous. Once a particular cluster was defined, the spike
each of the different data sets (tetrode, stereotrode, and train of each cell was computed by recovering the time
electrode) spike separation was achieved by an interactive stamp of each data point in the cluster.
computer program in which ellipses were drawn around
clusters of data points. Points within each cluster were
assigned a unique color and clusters could be iteratively 7. Results
followed and redefined in as many projections as needed,
or available, to uniquely define a particular single unit. For The performance of this identification procedure as
each parameter, tetrodes allowed for the display of 6 applied to tetrode, stereotrode and single electrode classifi-
possible 2-D projections. Comparisons between parameters cation is illustrated for a single recording site in Figs. I
allowed for an additional 6 projections for each pair and 2. Fig. 1 shows each of the six 2-D scatter plots of
yielding a total of 94 possible 2-D projections. Stereotrodes peak-to-peak amplitude obtained from a single tetrode that
provided 17 possible 2-D projections while single elec- yielded 6 single units. These plots reveal clearly separated
trodes allowed for 11 projections. In general we found that clusters reflecting the different amplitude ratios of the
plots of peak-to-peak amplitude provided the greatest re- different cells recorded on each of the 4 wires. The number
solving power for the tetrode and stereotrode data, while of well-identified clusters varies from one projection to
the principal components gave the best separation in the another and there often exist ambiguous boundaries be-
single electrode data. We therefore relied on these parame- tween adjacent clusters.
ters for spike separation and used the other parameters to These shortcomings are even more pronounced when
refine cluster boundaries when the identification was am- comparing the performance of tetrodes to that of each of

Table I
For each of 28 recording sites the number of isolated single units is listed for the tetrode, the best of the 6 possible stereotrode pairs. and the best of the 4
possible single electrodes
__.-_ -.----- ---_
Recordina site Tetrode Number of units isolated PCA PC.1 *
Stereotrode Stereotrode *
-..-- -..--..___.-
I 3 3 (3/6) 3 (l/6) 4 (l/4) 2 (I :‘J)
2 5 4 (2/6) 3 (l/6) 5 (2/4) 3 (Z,‘J)
i 6 6 (2/6) 6 (l/6) 5 (2/4) 5 (l/4)
.-1 5 2 (l/6) 2 (l/6) 4 (3/4) 4 (l/4)
5 4 3 (6/6) 2 (5/6) 3 (Z/4) i (i ,‘-I)
6 3 3 (4/6) 3 (4/6) 3 (3/4) 2 (L/4)
7 7 5 (2/6) 4 (l/6) 6 (l/4) .l (I~‘4
3 4 4 (l/6) 4 (l/6) 4 (l/4) ? (!,/4)
‘3 3 3 (4/6) 3 (l/6) 3 (4/4) I (i 1’4)
10 3 3 (2/6) 3 (3/6) 4 (l/4) ? (2A$)
11 4 4 (l/6) 4 (l/6) 4 (l/4) 2 i1./4)
12 6 4 (2/6) 4 (l/6) 5 (l/4) 3 i,I ‘4!
13 3 3 (5/6) 3 (5/6) 3 (4/4) I (2.‘4)
14 4 4 (2/6) 4 (l/6) 5 (2/4) 7 ij .‘J)
15 5 5 (S/6) 5 (5/6) 5 (l/4) 1 (2:‘J)
16 5 5 (l/6) 5 (i/6) 5 (3/4) ‘1 (1.J-t)
17 6 5 (l/6) 3 (3/6) 6 (l/4) 3 Cl/‘?)
18 I 5 (2/6) 3 (4/6) 6 (l/4) 4 li:‘4)
19 7 7 (l/6) 7 (l/6) 7 (l/4) h :, ,‘A)
20 7 7 (2/6) 7 (2/6) 6 (l/4) 4 !l,j4)
21 9 7 (l/6) 2 (l/6) 5 (l/4) &’ (2,‘4)
22 7 7 (I,‘61 5 (l/6) 4 (l/4) 5 iZ/‘4j
23 6 4 (3/6) 2 (l/6) 6 (l/4) 4 (I 4)
24 6 5 (l/6) 3 (2/6 5 (l/4) .:, ii,/41
25 6 4 (4/6) 3 (3/6) 6 (2/4) 4 !l/l)
26 5 3 (l/6) 3 (2/6) 5 (l/4) h l/4)
27 8 3 (2/6) 2 (3/6) 6 (l/4) 4 !i,,‘4)
28 7 4 (2/6) 4 (l/6) 7 (2/4) 4 / j ,:“$I

Total 151 122 (2.3/6) 102 (2/6) 137 (1.6/4) 9s ! 12,‘4f

Mean 5.4 4.4 3.6 4.9 3.4
.____-____
The ratios shown in parentheses indicate how many of the 6 stereotrode pairs or 4 electrodes were capable of isolating the number umts shown in each of
the 2 columns, respectively. Columns 4 and 6 show the number of cells isolated when the tenode classification is mapped on to the stereotrode and single
electrode data.
50 CM. Gray et at. /Journal of Neuroscience Methods 63 (I 995) 43-54

the 4 electrodes considered separately. Fig. 2 shows plots nel c, for example, the amplitudes and shapes of the
of the 1st versus 2nd principal components of the spike different spikes do not provide sufficient information for
waveforms computed for the data collected on each of the separation (compare A, B and C>. This is apparent in the
4 wires of the tetrode illustrated in Fig. 1. In these plots lack of well-separated clusters in Fig. 2 (i.e., electrode 3).
there are fewer non-overlapping clusters and the ambiguity Thus, these data demonstrate that the ratios of the spike
between cluster boundaries is greater. amplitudes of the different cells on the 4 channels provide
Following classification it is possible to observe how a more stringent criterion for identification than either the
the ambiguity and even overlap among clusters can be absolute amplitude or the waveform of the spikes on any
resolved (Fig. 1). What appear as fuzzy boundaries be- given channel.
tween adjacent clusters become clearly separable in a In order to appreciate the performance of the tetrode
different projection. In projection (3-4), the cluster en- recording technique Figs. 3 and 4 illustrate the post-stimu-
closed in an ellipse represents the combination of 2 differ- lus-time histograms (PSTH) of the cells recorded by the
ent cells (A,B) that can be resolved into two distinct tetrode shown in Figs. 1 and 2, and by a second tetrode,
clusters when utilizing the information available in projec- respectively. In both plots the isolated cells display clear
tion l-2. These findings illustrate that while stereotrodes, variations in firing patterns.
and single electrodes, may occasionally perform as well as The apparent superior performance of tetrodes >
tetrodes they can often lead to at least 2 types of classifica- stereotrodes > electrodes illustrated in Figs. 1 and 2 was
tion error: incomplete segregation of adjacent clusters and evaluated using our total sample of 28 recordings in 3
the lumping of 2 or more units into a single cluster. different animals. We adopted a 2-stage approach to this
The source of these ambiguities becomes apparent when analysis. First, we classified each data set according to the
observing the raw spike waveforms (Fig. IA-F). On chan- number of distinct clusters that could be identified. Here.

A B

24ms 234561
Spikes/Burst

C D

c
83
.-
8
#
I5 ,““““‘/““‘~“‘I”“““‘I
-+
0 4 6 8 IO .5 I 1.5 2
Time (ms) Time (ms)
Fig. 5. The shape and amplitude of action potentials produced by a cortical burst-firing neuron change during the course of a burst. A: plot of a short epoch
of raw data collected during the response to an optimal visual stimulus. Several bursts are displayed that show the characteristic changes in spike shape
which often consist of a decrease in amplitude and an increase in width. B: cumulative histograms of spike amplitude for each spike in a burst as a function
of the number of spikes that occur during a burst. The plot is normalized to the largest spike in the data set. Amplitude is largest during the first spike in a
burst and decreases thereafter. The magnitude of decrease is proportional to the number of spikes in the burst. C: interspike interval histogram of the spike
train (time resolution is 50 r~s). The clear refractory period in the distribution is evidence that only a single neuron is present in the recording. D: cluster
plot of peak-to-peak spike amplitude versus spike width for all spikes recorded in the data set. The distribution is unimodal but shows clear modulation of
both parameters. The range in spike amplitude varies by as much as 70% while spike width inversely co-varies over a range up to 40%.
 CM. Gray er al. /Journal of Neurosciencr Merftods 63 I I995 I 43-54 5;

we assumed that the number of clusters was equivalent to pairs (column 31, or electrodes (column 5) attaining the
the number of cells. The summary of this analysis is given best performance. These latter data indicate that the best
in columns 2, 3 and 5 of Table 1. Using these numbers performance was reached on average by 2.3 out of 6
alone we found that the tetrodes isolated 5.4 cells on possible stereotrode pairs, and by 1.6 out of 4 possible
average, while the best of 6 stereotrode pairs and the best electrodes. In addition, the best of 6 stereotrode pairs and
of 4 electrodes isolated 4.4 and 4.9 cells, respectively. The the best of 4 electrodes performed as well as the tetrode at
numbers in parentheses indicate the fraction of stereotrode 14 out of 28. and 17 out of 28 sites, respectively.

2-3 2-3

C D

Fig. 6. Tetrodes improve the isolation of bursting cell spike trains as compared to conventional single-electrode recording. A: cluster plot of peak-to-peak
amplitude of all spikes recorded on channels 2 and 3 of a single tetrode. Note the appearance of several elongated clusters. B: following the cluster cutting
procedure one of the isolated cells has a highly elongated cluster of points characteristic of the amplitude modulation displayed by bursting cells. The
cluster is linear in each of the 6 different 2-channel projections (data not shown). C: cluster plot of the 1st versus 2nd principal components of the same
data set shown in A. D: cluster plot of the principal components of the same cell as shown in B. Note that the cluster is not unimodal, indicating that the
spike waveforms vary in shape, thereby making the identification of the cluster boundaries diff&xlt to identify. E: plots of 10 overlapping waveforms from
each of the 4 channels of the tetrode sampled at 4 different locations in the elongated cluster shown in B. The waveforms on the left correspond to the
arrow on the right in B. Note that although the spikes displayed from different locations of the cluster differ markedly in amplitude the relative amplitude
ratios remain constant. This effect accounts for the linearity of the cluster shape and the ability of the tetrode technique to identify such cells.
52 CM. Gray et al. / Journal of Neuroscience Methods 63 11995) 43-54

For our second approach to the analysis, we mapped the The data collected from tetrodes and stereotrodes pro-
cluster assignments obtained from the tetrode analysis onto vides another source of information that is not available
both the stereotrode and principal components projections. from single electrode measurements. By comparing the
Thus, all the color assignments in the latter data sets were average amplitude difference among spikes to the mean
obtained solely from the ten-ode cluster cutting procedure. spatial separation between each of the wires of a tetrode it
This enabled us to determine the number of non-overlap- is possible to compute the decay of spike voltage as a
ping clusters having a single color, and if more than one function of distance (i.e., the ‘seeing distance’ of the
cluster had the same color, indicating that a single spike tetrode). To do this we calculated the average and standard
train had been incorrectly split into two. Using this ap- deviation of the difference in mean peak-to-peak voltage
proach, we found very different results (columns 4 and 6, for all the cells recorded on each of the 6 stereotrode pairs
Table 1). As verified by the tetrode classification, at all 28 recording sites. For any given cell and stereotrode
stereotrodes yielded a total of 102 cells, while single pair the peak-to-peak amplitude of the larger spike was
electrodes isolated a total of 9.5 cells. The best of the normalized to a value of 1. On average spike voltage
stereotrode pairs performed as well as the tetrode at 12 out decreased by a factor of 0.59 k 0.26 over a distance of 15
of 28 sites, while at no site did the principal components pm (n = 966).
analysis perform as well as the tetrode. The differences It can be assumed that under conditions of radial den-
between the two sets of results stem from 2 types of dritic symmetry and passive dendritic membrane that the
classification error: first, data points assigned by the ten-ode amplitude of the potential associated with somatic activa-
analysis to fall within a single cluster often overlapped, or tion decreases inversely with respect to the square of the
were mixed with, data points from another cluster; and distance from the cell’s soma (Rail, 1962):
second, clusters identified by the tetrode analysis often
split into two clusters when viewed using the principal
components projections. These errors are not detectable if
the user only has access to the stereotrode or electrode where V, is the voltage at distance r, Z, is the radial
data. Thus, if one takes into account the fact that fewer current, R, is the resistivity of the extracellular medium,
cells are isolated on average from a fraction of the and p is the cell radius. For cells of soma diameter of
stereotrode pairs and single electrodes, and that a signifi- lo-30 pm the voltage at a distance of 60 pm will be
cant number of classification errors take place, the proba- about 10% of the voltage value at the soma (Lemon,
bility that any given stereotrode or electrode penetration 1984). The radial decay of current can be modeled as an
will perform at the level of the average tetrode is markedly exponential relationship (Rail, 1962) leading to an equa-
reduced. tion for voltage decay having the general form of Eq. 5.
The tetrode and, to a lesser extent, the stereotrode From this equation one can compute the average voltage
recording techniques prevent a third source of classifica- decay as a function of distance from the soma. Thus, for
tion error that waveform recognition techniques are occa- any two channels of a tetrode (x, and x,> the voltage
sionally prone to make. In cortex and other structures, recorded from a distant soma will be:
burst-firing cells are often encountered in which both the
amplitude and time course of the action potentials change V( x,) = V(0) * ex”A (5)
during the burst (Ranck, 1973; Gray, 1992). The most
V( x2) = V(0) * ex2’* (6)
common pattern observed is a reduction in amplitude and
an increase in spike width in going from the first to the last where V(0) is the voltage at the point source and h is the
spike in a burst (Fig. 5). Waveform recognition techniques, decay constant. The measured decay rate of voltage is
such as template matching and principal components anal- equivalent to V(x,)/V(x,) = 0.59. Thus, the ratio be-
ysis, that are designed to be sensitive to these changes will tween Rqs. 5 and 6 is:
often identify more than one cell when only one is present 0.59 = e(%-x2vA
in the recording. This type of classification error may go (7)
unnoticed because the resulting spike trains will each Because the mean channel separation of the tetrodes is 15
continue to display a refractory period. When such cells pm the computed value for the decay constant is:
are encountered in a tetrode recording the resulting clusters h = 28.42 pm
are highly elongated but linear and most often continuous
in shape (Fig. 6). Moreover, the variation in spike ampli- From this value we can estimate the distance at which a
tude observed in one 2-D projection is similar to that seen 90% reduction in voltage from the point source will occur:
in each of the other projections (data not shown). Thus, 0.1 = e(x)/28.42w
because the ratio of spike amplitude remains relatively O-9
constant across channels, the tetrode technique is much where x = 65.6 pm. This value agrees with both empirical
less prone to the type of classification error introduced by (page 31 in Lemon, 1984) and theoretical (Fig. 10 in Rall,
changes in spike amplitude and shape. 1962) estimates of the decay of extracellular spike voltage.
 C.M. Gray et al. /Journal of Neuroscience Methods 63 f 1995) 43-54 5.3

These data indicate that the spike signals detected by a opportunity to evaluate other parameters of the spikes that
microelectrode are likely to arise, on average, from loca- might aid in separation (Kreiter et al., 1989). Here again
tions no further than 65 pm. At this distance, however, the the stereotrode and tetrode techniques provide additional
spikes are likely to be barely separable from the noise in information in the form of amplitude ratios across channels
the signal. Thus, on average, each channel of the tetrode to allow for the separation of clusters. Clusters that overlap
should be recording from a volume of tissue with a in one projection often are completely separable in another
diameter of approximately 130 pm, assuming uniform projection. Performance is improved by increasing the
spread of voltage in the volume of the cortical tissue. number of channels in the recording. Under ideal condi-
Therefore, this method is well suited to the study of small tions only one cell would produce a spike of equal ampli-
groups of neurons within a highly localized region of tude on each of the four channels of a tetrode and this
cortex. amplitude pattern would be different from all other ceils.
A third type of classification error was encountered
when we mapped the color assignments of the tetrode onto
8. Discussion the principal components projections. in a number of cases
we found spatially separated clusters having the same
The results of the present study demonstrate that tetrodes color. In these instances what appear to be two distinct
increase the yield and reliability of single-unit isolation clusters (cells) actually arise from variations in the wave-
from multi-unit recordings in the neocortex over that which form of a single cell. This type of classification error
can be obtained from stereotrodes and conventional micro- encountered with conventional recording techniques occurs
electrodes. The method achieves this improved perfor- when cells exhibit changes in their spike shape over time.
mance by reducing or eliminating several types of classifi- This occurs most often among cells that fire in rapid bursts
cation errors that are commonly encountered in single-unit ( Figs. 5 and 6) (Gray, 1992) and is thought to be due in
identification methods. The most common classification part to sodium channel inactivation. As a result the later
error for each of these methods is likely to arise from the spikes occurring within a burst are smaller in amplitude
incomplete segregation of clusters in the 2-D projection of and broader in width due to a decrease in the slope of the
any given pair of variables. As can be seen in Figs. 1 and 2 rising and falling phases of the action potential. If the
adjacent clusters may be clearly identifiable but ambiguous change in spike shape is sufficiently large, waveform
in the sense that the boundary between them is not explicit. recognition methods will incorrectly identify more than
Thus, both qualitative and quantitative methods employed one cell. As with the errors discussed above, the tetrode
to assign a boundary between such distributions will neces- technique avoids this problem by providing for a compari-
sarily result in a certain percentage of classification errors. son of amplitude ratios across channels. Thus, spike ampli-
The incidence of these errors will increase as a function of tude modulation is constant across chatmcls and the clus-
the overlap among the distributions. The important contri- tering of spike parameters remain contiguous across pro-
bution of the tetrode technique is that it provides an jections (Fig. 6). Even in those instances where the varia-
additional, and powerful, constraint (e.g., 6 pairs of ampli- tion of spike amplitude is sufficient to produce discontinu-
tude ratios) with which to separate overlapping clusters. ous clustering, the discontinuities are preserved across the
Thus, spike parameter distributions having overlapping different projections. Such covariation would not be ex-
clusters become clearly separable when the information pected to occur among different cells. We suspect that
available in the multiple projections of the tetrode is waveform variations of this, and other, sorts contributed to
employed. those instances in which the mapping of color assignments
A second, and potentially serious, classification error from the tetrode to the principal components projections
that can arise in multi-unit recordings is the tendency to resulted in spatially separate clusters having the same
group different cells with similar waveforms (amplitude color.
and time course) together into the same spike train. Such In addition to the benefits, there are at least two pitfalls
errors can result in ‘single-unit’ spike trains that actually associated with the methods employed in this study. First,
contain the activity of two or more cells. In many instances cluster separation and cell identification is fully dependent
such an error can probably be detected by the absence of a on the subjective criteria employed by the experimenter.
clear refractory period in the interspike interval distribu- To minimize error, we required that the cluster of a given
tion of the spike train. In such a case, however, the data cell be completely non-overlapping with all other clusters
must be discarded if one is solely interested in the proper- in at least one projection. In spite of this, what constitutes
ties of single rather than multi-unit activity. This type of a cluster and how the clusters are cut in the different
control may not be entirely foolproof, however, if the cells projections, will vary from user to user. Therefore, such
fire at low rates or if they are active at different times. supervised classification is likely to result in errors that are
Moreover, methods that are designed to separate single difficult to identify. Tbe spike trains that are isolated for a
units on-line are particularly vulnerable to such errors given recording are likely to be different in unpredictable
because once the classification has been made there is no ways from user to user. Clearly an unsupervised classifica-
54 C.M. Gray et al. /Journal of Neuroscience Methods 63 (1995) 43-54

tion method is needed to take full advantage of the power the deep layers of mouse somatosensory cortex. Neurosci. Len., 99:
inherent in the tetrode technique. Several unsupervised 137-141.
Connors, B.W., Gumick, M.J. and Prince, D.A. (1982) Electrophysiologi-
clustering methods have been developed and tested by cal properties of neocortical neurons in vitro. J. Neurophysiol., 48:
Zardosti and Wheeler (1992). The results of these studies 1302-1320.
indicate that the variability of classification is markedly Gray, C.M. (1992) Bursting cells in visual cortex exhibit properties
reduced in comparison to subjective methods. characteristic of intrinsically bursting cells in sensorimotor cortex.
The second pitfall of the methods employed here, and Sot. Neurosci. Abstr., 18: 131.2.
Gray, C., Engel, A.K., Koenig, P. and Singer, W. (1990) Stimulus-depen-
one which applies to the vast majority of spike separation dent neuronal oscillations in cat visual cortex: receptive field proper-
techniques, concerns the spike overlap problem. Two cells ties and feature dependence. Eur. J. Neurosci., 2: 607-619.
that fire at the same time will produce a complex superpo- Jansen, R.F. and Ter Maat, A. (1992) Automatic wave form classification
sition of voltage having a waveform and amplitude distri- of extracellular multineuron recordings. J. Neurosci. Methods, 42:
bution across the channels that most often falls outside the 123-132.
Kreiter, A.K., Aertsen, A.M.H.J. and Gerstein, G.L. (1989) A low-cost
boundaries of any one cluster. This problem is com- single-board solution for real-time unsupervised waveform classiflca-
pounded by variations in the phase of the overlap and by tion of multineuron recordings. J. Neurosci. Methods, 30: 59-69.
the combinatorial nature of the overlap when there are Lemon, R. (1984) Methods for neuronal recording in conscious animals.
more than 2 cells active at the same time. Moreover, in IBRO Handbook Series: Methods in the Neurosciences, Vol. 4, John
structures such as visual cortex that exhibit high firing Wiley, New York.
Lewicki, M. (1994) Bayesian modeling and classification of neural
rates and synchronized activity, this problem is com- signals. Neural Comp., 6: 1005-1030.
pounded even further. In effect, the clustering methodol- McCormick, D.A., Connors, B.W., Lighthall, J.W. and Prince, D.A.
ogy becomes an efficient notch filter for overlapping spikes. (1985) Comparative electrophysiology of pyramidal and sparsely
This problem is faced by each of the methods employed to spiny stellate neurons of the neocortex. 1. Neurophysiol., 54: 782-806.
separate spikes. Thus, until effective algorithms can be McNaughton, B.L., O’Keefe, J.O., Barnes, C.A. (1983) The stereotrode:
a new technique for simultaneous isolation of several single units in
implemented to recover overlapping spikes (Prochazka and the central nervous system from multiple unit records. I. Neurosci.
Komhuber, 1973; Lewicki, 1994) a certain fraction of the Methods 8: 391-397
data in a spike train will routinely be lost. Tetrodes may O’Keefe, J.O. and Reece, M.L. (1993) Phase relationship between hip-
have an additional advantage in constraining the search for pocampal place units and the EEG theta rhythm. Hippocampus, 3:
overlapping spikes. 317-330.
Pettigrew, J.D., Cooper, M.L. and Blasdel, G.G. (1979) Improved use of
tapetal reflection for eye-position monitoring. Assoc. Res. Vis. Oph-
dial., 18: 490-495.
Acknowledgements Prochazka, V.J. and Komhuber, H.H. (1973) On line multiunit sorting
with resolution of superposition potentials. Electroenceph. Clin. Neu-
We thank Bruce Wheeler, Ken Miller, and George rophysiol., 34: 91-93.
Rall, W. (1962) Electrophysiology of a dendritic neuron model. Biophys.
Gerstein for their helpful comments on an earlier version
I., 2: 145-167.
of the manuscript. We also thank George Gerstein’s contri- Ranck, J.B. Jr. (19731 Studies on single neurons in dorsal hippocampal
bution to the calculations of principal components. Hi- formation and septum in unrestrained rats. Exp. Neural., 41: 461-555.
royuki Ito’s contribution to the calculations of the spatial Salganicoff, M. Sama, M., Sax, L., Gerstein, G.L. (1988) Unsupervised
decay of voltage were greatly appreciated. This ‘work was waveform classification for multi-neuron recordings: a real-time, soft-
ware-based system. I. Algorithms and implementation. J. Neurosci.
supported by grants from The Office of Naval Research Methods, 12: 181-187
(N00014-91-J-12561, The National Eye Institute Wheeler, B.C. and Heetderks, W.J. (1982) A comparison of techniques
(EYO86861, and the Klingenstein Foundation to C.M.G., for classification of multiple neural signals. IEEE Trans. Biomed.
and the National Institute of Mental Health (MH46823) to Eng., 29: 752-759.
B.M. Wilson, M.A. and McNaughton, B.L. (1993) Dynamics of the hippocam-
pal ensemble code for space. Science, 261: 1055-1058.
wiirgijtter, F., Daunicht, W. and E&miller, R. (1986) An on-line spike
form discriminator for extracellular recording based on an analog
References correlation technique. J. Neurosci. Methods, 17: 141-151.
Zardosti, M. and Wheeler, B. (1992) Finding the number of neurons in
Abeles, M. and Goldstein, M.H. (1977) Multispike Train Analysis. Proc. extracellular recordings: a comparison of the effectiveness of popular
IEEE, 65: 762-773. clustering indices. In: M.A. Penna (Ed.), hoc. Computational Neuro-
Agmon, A. and Connors, B.W. (1989) Repetitive burst-firing neurons in science Symposium, pp. 229-235, October, Indianapolis.